Simplicity is a great virtue but it requires hard work to achieve it and education to appreciate it. And to make matters worse: complexity sells better. - Edsger Dijkstra ASYMPTOTIC COMPLEXITY Lecture 10 CS2110 Spring 2019
Announcements 2 Next Mon-Tues: Spring Break No recitation next week Regrade requests will be processed this weekend Prelim is on Tuesday, 12 March. Prelim Review for prelim Sunday, 10 March, 1-3PM Next Thursday, we will tell you What time you will be assigned to take it What to do if you can t take it then but can take the other one What to do if you can t take it that evening. What to do if authorized for more time or quiet space
Help in providing code coverage 3 White-box testing: make sure each part of program is exercised in at least one test case. Called code coverage. Eclipse has a tool for showing you how good your code coverage is! Use it on A3 (and any programs you write) JavaHyperText entry: We demo it. code coverage
What Makes a Good Algorithm? 4 Suppose you have two possible algorithms that do the same thing; which is better? What do we mean by better? Faster? Less space? Simpler? Easier to code? Easier to maintain? Required for homework? FIRST, Aim for simplicity, ease of understanding, correctness. SECOND, Worry about efficiency only when it is needed. How do we measure speed of an algorithm?
Basic Step: one constant time operation 5 Constant time operation: its time doesn t depend on the size or length of anything. Always roughly the same. Time is bounded above by some number Basic step: Input/output of a number Access value of primitive-type variable, array element, or object field assign to variable, array element, or object field *** do one arithmetic or logical operation method call (not counting arg evaluation and execution of method body)
Counting Steps 6 // Store sum of 1..n in sum sum= 0; // inv: sum = sum of 1..(k-1) for (int k= 1; k <= n; k= k+1){ sum= sum + k; Statement: # times done sum= 0; 1 k= 1; 1 k <= n n+1 k= k+1; n sum= sum + k; n Total steps: 3n + 3 All basic steps take time 1. There are n loop iterations. Therefore, takes time proportional to n. 350 300 250 200 150 100 50 0 Linear algorithm in n 0 20 40 60 80 100
Not all operations are basic steps 7 // Store n copies of c in s s= ""; // inv: s contains k-1 copies of c for (int k= 1; k <= n; k= k+1){ s= s + 'c'; Statement: # times done s= ""; 1 k= 1; 1 k <= n n+1 k= k+1; n s= s + 'c'; n Total steps: 3n + 3 Catenation is not a basic step. For each k, catenation creates and fills k array elements.
String Catenation 8 s= s + c ; is NOT constant time. It takes time proportional to 1 + length of s String@00 b char[] String s String@90 b char[] String char[]@02 0 d 1 x char[] char[]@018 0 d 1 x 2 c char[]
Not all operations are basic steps 9 // Store n copies of c in s s= ""; // inv: s contains k-1 copies of c for (int k= 1; k <= n; k= k+1){ s= s + 'c'; Statement: # times # times # steps done s= ""; 1 1 1 k= 1; 1 1 1 k <= n n+1 n+11 k= k+1; n n 1 s= s + 'c'; n n k Total steps: n*(n-1)/2 3n + + 3 2n + 3 Catenation is not a basic step. For each k, catenation creates and fills k array elements. 350 300 250 200 150 100 50 0 Quadratic algorithm in n 0 20 40 60 80 100
Linear versus quadractic 10 // Store sum of 1..n in sum sum= 0; // inv: sum = sum of 1..(k-1) for (int k= 1; k <= n; k= k+1) sum= sum + n Linear algorithm // Store n copies of c in s s= ; // inv: s contains k-1 copies of c for (int k= 1; k = n; k= k+1) s= s + c ; Quadratic algorithm In comparing the runtimes of these algorithms, the exact number of basic steps is not important. What s important is that One is linear in n takes time proportional to n One is quadratic in n takes time proportional to n 2
Looking at execution speed 11 Number of operations executed 2n+2, n+2, n are all linear in n, proportional to n n*n ops 2n + 2 ops n + 2 ops n ops Constant time 0 1 2 3 size n of the array
What do we want from a definition of runtime complexity? 12 Number of operations executed 0 1 2 3 n*n ops 2+n ops 5 ops size n of problem 1. Distinguish among cases for large n, not small n 2. Distinguish among important cases, like n*n basic operations n basic operations log n basic operations 5 basic operations 3. Don t distinguish among trivially different cases. 5 or 50 operations n, n+2, or 4n operations
"Big O" Notation 13 Formal definition: f(n) is O(g(n)) if there exist constants c > 0 and N 0 such that for all n N, f(n) c g(n) c g(n) f(n) Get out far enough (for n N) f(n) is at most c g(n) Intuitively, f(n) is O(g(n)) means that f(n) grows like g(n) or slower N
Prove that (2n 2 + n) is O(n 2 ) 14 Formal definition: f(n) is O(g(n)) if there exist constants c > 0 and N 0 such that for all n N, f(n) c g(n) Example: Prove that (2n 2 + n) is O(n 2 ) Methodology: Start with f(n) and slowly transform into c g(n): Use = and <= and < steps At appropriate point, can choose N to help calculation At appropriate point, can choose c to help calculation
Prove that (2n 2 + n) is O(n 2 ) 15 Formal definition: f(n) is O(g(n)) if there exist constants c > 0 and N 0 such that for all n N, f(n) c g(n) Example: Prove that (2n 2 + n) is O(n 2 ) f(n) = <definition of f(n)> 2n 2 + n <= <for n 1, n n 2 > 2n 2 + n 2 = <arith> 3*n 2 = <definition of g(n) = n 2 > 3*g(n) Transform f(n) into c g(n): Use =, <=, < steps Choose N to help calc. Choose c to help calc Choose N = 1 and c = 3
Prove that 100 n + log n is O(n) 16 Formal definition: f(n) is O(g(n)) if there exist constants c > 0 and N 0 such that for all n N, f(n) c g(n) f(n) = <put in what f(n) is> 100 n + log n <= <We know log n n for n 1> 100 n + n = <arith> 101 n = <g(n) = n> 101 g(n) Choose N = 1 and c = 101
O( ) Examples 17 Let f(n) = 3n 2 + 6n 7 f(n) is O(n 2 ) f(n) is O(n 3 ) f(n) is O(n 4 ) p(n) = 4 n log n + 34 n 89 p(n) is O(n log n) p(n) is O(n 2 ) h(n) = 20 2 n + 40n h(n) is O(2 n ) a(n) = 34 a(n) is O(1) Only the leading term (the term that grows most rapidly) matters If it s O(n 2 ), it s also O(n 3 ) etc! However, we always use the smallest one
Do NOT say or write f(n) = O(g(n)) 18 Formal definition: f(n) is O(g(n)) if there exist constants c > 0 and N 0 such that for all n N, f(n) c g(n) f(n) = O(g(n)) is simply WRONG. Mathematically, it is a disaster. You see it sometimes, even in textbooks. Don t read such things. Here s an example to show what happens when we use = this way. We know that n+2 is O(n) and n+3 is O(n). Suppose we use = n+2 = O(n) n+3 = O(n) But then, by transitivity of equality, we have n+2 = n+3. We have proved something that is false. Not good.
Problem-size examples 19 Suppose a computer can execute 1000 operations per second; how large a problem can we solve? operations 1 second 1 minute 1 hour n 1000 60,000 3,600,000 n log n 140 4893 200,000 n 2 31 244 1897 3n 2 18 144 1096 n 3 10 39 153 2 n 9 15 21
Commonly Seen Time Bounds 20 O(1) constant excellent O(log n) logarithmic excellent O(n) linear good O(n log n) n log n pretty good O(n 2 ) quadratic maybe OK O(n 3 ) cubic maybe OK O(2 n ) exponential too slow
Search for v in b[0..] 21 Q: v is in array b Store in i the index of the first occurrence of v in b: R: v is not in b[0..i-1] and b[i] = v. Methodology: 1. Define pre and post conditions. 2. Draw the invariant as a combination of pre and post. 3. Develop loop using 4 loopy questions. Practice doing this!
Search for v in b[0..] 22 Q: v is in array b Store in i the index of the first occurrence of v in b: R: v is not in b[0..i-1] and b[i] = v. pre:b post: b inv: b 0 b.length v in here 0 i b.length " "? 0 i b.length " v in here Methodology: 1. Define pre and post conditions. 2. Draw the invariant as a combination of pre and post. 3. Develop loop using 4 loopy questions. Practice doing this!
The Four Loopy Questions 23 Does it start right? Is {Q init {P true? Does it continue right? Is {P && B S {P true? Does it end right? Is P &&!B => R true? Will it get to the end? Does it make progress toward termination?
24 pre:b post: b Q: v is in array b Search for v in b[0..] Store in i the index of the first occurrence of v in b: R: v is not in b[0..i-1] and b[i] = v. 0 b.length v in here 0 i b.length " "? i= 0; while ( b[i]!= v) { i= i+1; inv: b 0 i b.length " v in here Linear algorithm: O(b.length) Each iteration takes constant time. Worst case: b.length iterations
Binary search for v in sorted b[0..] 25 // b is sorted. Store in i a value to truthify R: // b[0..i] <= v < b[i+1..] pre:b post: b inv: b 0 b.length sorted 0 i b.length " > " 0 i k b.length "? > " b is sorted. We know that. To avoid clutter, don t write in it invariant Methodology: 1. Define pre and post conditions. 2. Draw the invariant as a combination of pre and post. 3. Develop loop using 4 loopy questions. Practice doing this!
Binary search for v in sorted b[0..] 26 // b is sorted. Store in i a value to truthify R: // b[0..i] <= v < b[i+1..] pre:b post: b inv: b 0 b.length sorted 0 i b.length " > " 0 i k b.length "? > " 0 i e k " " > " i= -1; k= b.length; while ( i+1< k) { int e=(i+k)/2; // -1 i < e < k b.length if (b[e] <= v) i= e; else k= e;
Binary search for v in sorted b[0..] 27 // b is sorted. Store in i a value to truthify R: // b[0..i] <= v < b[i+1..] pre:b post: b inv: b 0 b.length sorted 0 i b.length " > " 0 i k b.length "? > " Logarithmic: O(log(b.length)) i= -1; k= b.length; while ( i+1< k) { int e=(i+k)/2; // -1 e < k b.length if (b[e] <= v) i= e; else k= e; Each iteration takes constant time. Worst case: log(b.length) iterations
Binary search for v in sorted b[0..] 28 // b is sorted. Store in i a value to truthify R: // b[0..i] <= v < b[i+1..] This algorithm is better than binary i= -1; searches that stop when v is found. k= b.length; while ( i+1< k) { 1. Gives good info when v not in b. int e=(i+k)/2; 2. Works when b is empty. // -1 e < k b.length 3. Finds first occurrence of v, not if (b[e] <= v) i= e; arbitrary one. else k= e; 4. Correctness, including making progress, easily seen using invariant Each iteration takes constant time. Logarithmic: O(log(b.length)) Worst case: log(b.length) iterations
29 Dutch National Flag Algorithm
Dutch National Flag Algorithm Dutch national flag. Swap b[0..n-1] to put the reds first, then the whites, then the blues. That is, given precondition Q, swap values of b[0.n-1] to truthify postcondition R: Q: b R: b P1: b P2: b 0 n? 0 n reds whites blues 0 n reds whites blues? 0 n reds whites? blues Suppose we use invariant P1. What does the repetend do? 2 swaps to get a red in place
Dutch National Flag Algorithm Dutch national flag. Swap b[0..n-1] to put the reds first, then the whites, then the blues. That is, given precondition Q, swap values of b[0.n-1] to truthify postcondition R: Q: b R: b P1: b P2: b 0 n? 0 n reds whites blues 0 n reds whites blues? 0 n reds whites? blues Suppose we use invariant P2. What does the repetend do? At most one swap per iteration Compare algorithms without writing code!
Dutch National Flag Algorithm: invariant P1 Q: b R: b 0 n? 0 n reds whites blues 0 h k p n P1: b reds whites blues? h= 0; k= h; p= k; while ( p!= n ) { if (b[p] blue) p= p+1; else if (b[p] white) { swap b[p], b[k]; p= p+1; k= k+1; else { // b[p] red swap b[p], b[h]; swap b[p], b[k]; p= p+1; h=h+1; k= k+1;
Dutch National Flag Algorithm: invariant P2 Q: b R: b 0 n? 0 n reds whites blues 0 h k p n P2: b reds whites? blues h= 0; k= h; p= n; while ( k!= p ) { if (b[k] white) k= k+1; else if (b[k] blue) { p= p-1; swap b[k], b[p]; else { // b[k] is red swap b[k], b[h]; h= h+1; k= k+1; 33
Asymptotically, which algorithm is faster? 34 Invariant 1 Invariant 2 0 h k p n reds whites blues? h= 0; k= h; p= k; while ( p!= n ) { if (b[p] blue) p= p+1; else if (b[p] white) { swap b[p], b[k]; p= p+1; k= k+1; else { // b[p] red swap b[p], b[h]; swap b[p], b[k]; p= p+1; h=h+1; k= k+1; 0 h k p n reds whites? blues h= 0; k= h; p= n; while ( k!= p ) { if (b[k] white) k= k+1; else if (b[k] blue) { p= p-1; swap b[k], b[p]; else { // b[k] is red swap b[k], b[h]; h= h+1; k= k+1;
Asymptotically, which algorithm is faster? 35 Invariant 1 Invariant 2 0 h k p n reds whites blues? 0 h k p n reds whites? blues h= 0; k= h; p= k; while ( p!= n ) { might use 2 swaps per iteration if (b[p] blue) p= p+1; else if (b[p] white) { swap b[p], b[k]; p= p+1; k= k+1; else { // b[p] red swap b[p], b[h]; swap b[p], b[k]; p= p+1; h=h+1; k= k+1; h= 0; k= h; p= n; while ( k!= p ) { uses at most 1 swap per iteration if (b[k] white) k= k+1; else if (b[k] blue) { p= p-1; swap b[k], b[p]; else { // b[k] is red swap b[k], b[h]; h= h+1; k= k+1; These two algorithms have the same asymptotic running time (both are O(n))